Handling caches is a challenge. When the system is not under memory pressure, depending upon the memory v cache footprint of the application, heuristics would need to be designed to decide on the right time to power off memory. The overhead would be to write out dirty pages from the cache before actually turning off the memory. The regions framework would enable doing this targeted reclaim of areas of memory that need to be evacuated.
Posted Jun 10, 2011 13:40 UTC (Fri) by ejr (subscriber, #51652)
[Link]
Thank you.
Unfortunately, there is always memory pressure in my areas (massive graph analysis, numerics, etc.). I'm working on a memory concurrency v. performance model for some graph tasks with a thought towards powering down unneeded cores (think SCC) when they cannot contribute. Now I'm wondering if there's some way to consider the memory side, again assuming everything fits (which it doesn't).
There also is much, *much* work going into dropping NAND flash into DRAM slots (phase change, plus DRAM cache). That will change the power usage characteristics drastically. If you can turn off the DRAM cache without having to flush out all the data...
Memory power management
Posted Jun 11, 2011 0:29 UTC (Sat) by giraffedata (subscriber, #1954)
[Link]
I think it's usually the case that essentially all the memory is "used," so I don't get how the proposed policy can be effective. Memory is "used" to varying degrees, i.e. how important the contents are. Some is actually indispensable because there's no practical way to recreate its contents. But other memory is a cache of file contents, cache of dentries, memory that could be moved to swap space, and the like.
There's also internal fragmentation in memory allocation pools -- memory that's used just to anticipate future allocations and save time.
So I don't see any policy that just says power down totally free memory as being terribly useful. We need a policy that weighs for each page the value of having that data in memory vs the cost of powering that page.
Memory power management
Posted Jun 11, 2011 2:21 UTC (Sat) by Ankita (subscriber, #39147)
[Link]
Ah yes, the proposal aims to conserve memory power when the system is not under too much memory pressure. We are targeting scenarios where minimal consolidation of references/allocations can result in significant power savings.